DAC: Straight Talk with Lip-Bu Tan

For the last few DACs, every lunchtime the DAC Pavilion has an interview by Ed Sperling with an EDA CEO. These used to be called "fireside chats" but since the time a couple of years ago when Lip-Bu's chat was delayed by a real fire (and evacuation) they are now called "straight talk". This year, on the Tuesday of DAC, it was once again Lip-Bu Tan's turn.

"You cover a broad swath of the world," Ed opened, "as well as many markets. Which are the most important?"

Lip-Bu thought the data-driven economy. There are lots of devices and sensors, increasingly moving to the intelligent edge, to process at low power, and then scale to the cloud.

Last week I went to see Morris [Chang] when he retired, and I told him his biggeest contribution is that fabless companies don't need to build a fab. It's the same with Google and Microsoft massively scaling and changing the infrastrucutre requirements. We are also moving to the "vertical cloud". Mobile cloud, automotive cloud, and so on. We at Cadence see tremendous design activity going on. The semiconductor industry grew over 20% last year past $400B. In the next 5 years it will grow more than double the rate of the last 5 years.

Ed asked if the cloud was just at the beginning of its growth phase. Lip-Bu agreed it was, and now we are moving into domain-specific computing. "If you love innovation and changing the world, this is the best time."

Ed was worried about all the companies developing AI chips. "There are way to many companies. that can't all last."

Lip-Bu said AI and ML are broadly impacting the whole industry. "General purpose CPU is going to decline the next few years, replaced with GPU and ASIC. Workloads are changing and it's a paradigm shift. Genomic sequencing, automotive, medical. All are about big data. Any new paradigm shift gets a lot of startups with different points and over time the 100 or 200 end up as 2 or 3 that are world class, Of course, lots acquired along the way. It's an exciting time."

Ed wondered how flexibility gets built in. "It's not like an ASIC for smartphone, sometimes it seems to change in weeks or days."

Lip-Bu agreed:

This is where the excitement comes in. We have inside and outside approach. We have machine learning inside the tools to shorten the design cycle and verification needs. Other part is to optimize our tools for some specific applications talking about algorithm compiler for the silicon you want to optimize. A GPU is great for some things but really power hungry. If it is industrial IoT it is neuromorphic: wake up, process, go back to sleep with no latency. But a lot depends on what application you are trying to drive.

Ed switched gears to advanced nodes. Lip-Bu said that we are in 7nm with many customers, engaged at 5nm, and are doing test chips at 3nm. It's how we drive the power and cost down. "Our job is to help our customers design the most sophisticated silicon. It’s become more complex, it's not just digital it’s analog mixed signal. Also, the packaging and power and signal integrity. We need to look at the totality, look at the system point of view."

Ed wondered who is going to win the automotive wars as we move to ADAS and autonomous driving. "Do you see the industry power changing? Who will run this, the car makers or Apple/Google?"

Lip-Bu makes no secret of loving his Tesla that drives itself much of the way to Cadence so he can make calls safely at the same time. But as to industry structure, "It is clearly a race, everyone from tier-1 to OEMs to people coming from the IT side. We have a long way to go. Security is a big issue. All the sensors—Lidar, Radar, and cameras—mean you need a supercomputer in the trunk and we need to reduce in size and make more cost effective. GM is really aggressive with Cruze automation, and they are making a lot of progress. It will happen in our lifetime."

Ed wondered about electrification of automobile traction and what Lip-Bu thinks will happen.

It is hard to predict. Both electric and gas for a time. Charge stations are in the Bay Area everywhere, but in other areas not so much. I have a hybrid car too, and I can see if I can make it with electric and charging stations, and if not take the other car.

"What about AI in EDA?" Ed asked. "Will tools be able to improve themselves?"

Yes, this is not a 1-year thing, it is just starting with place and route, and now starting to apply to verification, and also custom/analog with Virtuoso. In the end, we are serving our customers and we need to make them win (and then they buy more of our tools).

Another aspect of AI that Ed was worried about was privacy.

Security and privacy is something our industry needs to address. Right now it is all about data, medical data, household data, credit cards, If you are driving, you generate a lot of data. But more concerning is data being compromised on the business side. Putting design data in the cloud is even more sensitive. But we have worked closely with foundry and IP and cloud partners to make sure it is properly protected. We have very strong CIO and his team, need to make sure our own data is secure, and also that of our customers in the hyperscale cloud.

Ed went back to advanced nodes and the struggle to move down Moore's Law. He is worried whether designs can be done in a reasonable amount of time, and whether there are enough customers.

Well, for automotive and IoT we don’t need to go down immediately, we can use 90nm, 28nm for a lot of designs, especially analog. On digital side we are marching down to 5nm, 7nm is in production. When you move down to 5nm and 3nm, then you need EUV and multiple patterning. We already announced a 3nm test chip with imec. We need to be out ahead of our customers. Our team is stretched to drive the leading edge process nodes. Next year will have 5nm chips. There is still lots of room to go on the process roadmap.  The best way is for us is to work with the best customers so we know what they need, then the rest will follow. Our R&D is highest in the industry because we are thinking 5-10 years ahead.

"What aboout IP?" Ed wondered.

Lip-Bu pointed out that IP is a good business, and last year Cadence's IP revenues grew 18%. One driver is that outsourcing is becoming more and more acceptable to customers since it is a nightmare to match all the protocols. One difference, Lip-Bu pointed out, is that a bug in an EDA tool you can work around (usually), but a bug in IP gets designed in. "We are focused on mission-critical IP, such as Tensilica and Nusemi. Over time the customer comes to us for silicon proven IP in the most advanced nodes. My IP team is focused on 7nm, 5nm and 3nm. We need to be there for our customers."

Ed's last question was what Cadence would look like several years out.

Lip-Bu pointed out that we are putting a lot of effort into some vertical markets.

To grow the pie you go up or you go horizontal, but we are doing both. Some verticals like mil aero, defense are important. Automotive we mentioned earlier. Datacenter guys. But there is a sea-change going on since each cloud supplier is building some portion of their silicon to scale out storage and network and datacenter. We want to be the EDA/IP partner of choice for them. Automotive is the same. We want to make the tool and EDA portfolio to be the partner of choice for them. So I think you will see us want to be best in class for EDA, then everything up to system modeling. Growing from generic systems to verticals. At end of the day we serve our customers to ensure they can build the best product in the fastest time, and right first time.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.